Probabilistic deep learning framework for dynamic carbon emission accounting of electric buses under grid uncertainty

The electrification of public transit has emerged as a pivotal pathway for deep urban decarbonization. However, existing carbon accounting methods predominantly rely on static grid emission factors and deterministic energy models, often overlooking spatiotemporal variability and uncertainty propagation. To address this limitation, this study establishes a dynamic, uncertainty-aware framework for the carbon accounting of electric bus systems. A hybrid deep learning architecture integrating Temporal Convolutional Networks (TCN), Bidirectional Long Short-Term Memory (BiLSTM) networks, and an Attention mechanism is developed to capture multi-scale temporal dependencies in energy consumption. In parallel, a time-varying probabilistic grid emission model is formulated using period-specific distributions across diurnal intervals, and Monte Carlo simulation is employed to propagate uncertainty throughout the accounting chain. Using high-frequency operational telemetry from ten electric buses in Shenzhen, the proposed model achieved an [Formula: see text] of 0.9610 and an RMSE of 0.0523, outperforming ensemble learning methods, conventional deep learning baselines, and ablation variants. Leave-one-bus-out cross-validation further confirmed robust cross-vehicle generalizability, with a mean [Formula: see text] of [Formula: see text]. The results reveal pronounced heteroscedasticity in carbon emission profiles, with uncertainty expanding substantially during high-power transient events, while the time-varying emission factor model yields wider confidence intervals than static approaches. These findings demonstrate the importance of uncertainty-aware dynamic accounting and provide a robust data-driven basis for probabilistic carbon footprint estimation in urban public transport.